2010 Lecture 2

Computational Fluid Dynamics I

http://users.wpi.edu/~gretar/me612.html! Objectives:!
Introduce the basic concepts needed to solve a partial

Elementary! differential equation using finite difference methods. !
Numerical Analysis:! Discuss basic time integration methods, ordinary and

partial differential equations, finite difference
Finite Difference approximations, accuracy. !
Approximations-II! Show the implementation of numerical algorithms into

actual computer codes.!
Grétar Tryggvason!
Spring 2010!


Outline!
•  Solving partial differential equations!

! !Finite difference approximations!
! !The linear advection-diffusion equation!
! !Matlab code!
! !Accuracy and error quantification!
Finite Difference
! !Stability!
! !Consistency!
Approximations of
! !Multidimensional problems!
! !Steady state! the Derivatives!


Finite Difference Approximations! Finite Difference Approximations!
Derive a numerical approximation to the governing
!
equation, replacing a relation between the derivatives by
a relation between the discrete nodal values.
The!
Δt

f (t + 2Δt) Time Derivative!
f (t + Δt)
Δt
f(t,x-h) f(t,x) f(t,x+h) !
f (t)
h
h


Time derivative!
The Time Derivative is found using a FORWARD
using a Taylor series !

EULER method. The approximation can be found by
f (t + Δt) = f (t) +
∂f (t)
∂t
Δt +
∂ 2 f (t) Δt 2
∂t 2 2
+
Solving this equation for the time derivative gives:!

f(t+Δt,x)

Δt
f(t,x)! ∂f (t) f (t + Δt) − f (t) ∂ 2 f (t) Δt
= − +
h
h
∂t Δt ∂t 2 2


Apply this to the ODE from last lecture:!
The original equation at time level n is:!

⎛ df ⎞
n
Solving a!
⎝ dt ⎠ = g
n
And using the approximation just derived for the partial differential
time derivative results in:!
f n+1 − f n
= g n + O(Δt)
equation!
Δt
which is exactly the first order forward Euler method. !


Model Equations! Model Equations!
For initial conditions of the form:!

We will use the model equation:!
f (x, t = 0) = Asin(2πkx)
∂f ∂f ∂2 f
+U =D 2 It can be verified by direct substitution that the solution
∂t ∂x ∂x is given by:!
f (x, t) = e − Dk t sin(2πk(x − Ut))
2
to demonstrate how to solve a partial equation numerically.!
Although this equation is much simpler than the full

Navier Stokes equations, it has both an advection term which is a decaying
and a diffusion term. ! traveling wave!
Before attempting to solve the equation, it is useful to
understand how the analytical solution behaves.!

When using FINITE DIFFERENCE approximations,

the values of f are stored at discrete points.!
The Spatial! f(x-h) f(x) f(x+h) !
First Derivative! h
h

The derivatives of the function are approximated

using a Taylor series!


Start by expressing the value of f(x+h) and f(x-h) in terms of f(x)

The result is:!
∂f (x) ∂ 3 f (x) h 3
∂f (x) ∂ 2 f (x) h 2 ∂ 3 f (x) h 3 ∂ 4 f (x) h 4 f (x + h) − f (x − h) = 2 h+2 +
f (x + h) = f (x) + h+ + + + ∂x ∂x 3 6
∂x ∂x 2 2 ∂x 3 6 ∂x 4 24
∂f (x) ∂ 2 f (x) h 2 ∂ 3 f (x) h 3 ∂ 4 f (x) h 4 Rearranging this equation to isolate the first derivative:!
f (x − h) = f (x) − h+ − + −
∂x ∂x 2 2 ∂x 3 6 ∂x 4 24
Subtracting the second equation from the first:! ∂f (x) f (x + h) − f (x − h) ∂ 3 f (x) h 2

= − +
∂x 2h ∂x 3 6
∂f (x) ∂ 3 f (x) h 3
f (x + h) − f (x − h) = 2 h+2 +
∂x ∂x 3 6


Start by expressing the value of f(x+h) and f(x-h) in terms of f(x)

∂f (x) ∂ 2 f (x) h 2 ∂ 3 f (x) h 3 ∂ 4 f (x) h 4

f (x + h) = f (x) + h+ + + +
The Spatial! ∂x ∂x 2 2 ∂x 3 6 ∂x 4 24
∂f (x) ∂ 2 f (x) h 2 ∂ 3 f (x) h 3 ∂ 4 f (x) h 4
Second Derivative! f (x − h) = f (x) −
∂x
h+
∂x 2 2
−
∂x 3 6
+
∂x 4 24
−
Adding the second equation to the first:!
∂f 2 (x) h 2 ∂ 4 f (x) h 4
f (x + h) + f (x − h) = 2 f (x) + 2 +2 +
∂x 2
2 ∂x 4 24

The result is:!
∂f 2 (x) h 2 ∂ 4 f (x) h 4
f (x + h) + f (x − h) = 2 f (x) + 2 +2 +
∂x 2
2 ∂x 4 24 Solving the partial
Rearranging this equation to isolate the second derivative:!
differential equation!
∂ 2 f (x) f (x + h) − 2 f (x) + f (x − h) ∂ 4 f (x) h 2
= − +
∂x 2 h2 ∂x 4 12


We already introduced the notation! f = f (t) A numerical approximation to!

n
For space and time we will use:! ∂f ∂f ∂2 f

+U =D 2
∂t ∂x ∂x
f jn = f (t, x j ) Is found by replacing the derivatives by the following
f jn+1 approximations!
f jn+1 = f (t + Δt, x j ) ⎛∂ 2 f ⎞
n
⎛ ∂f ⎞ ⎛ ∂f ⎞
n n
f jn−1 f jn f jn+1 Δt +U = D⎜ 2
f jn+1 = f (t, x j + h) ⎝ ∂t ⎠ j ⎝ ∂x ⎠ j ⎝ ∂x ⎠ j
f jn−1 = f (t, x j − h)
h
h

f jn+1
f jn−1 f jn f jn+1 Δt
h
h



Using our shorthand ! Substituting these approximations into:!

gives!
notation! ∂f ∂f ∂2 f
⎛ ∂f ⎞
n
f n+1 − f jn +U =D 2
= j + O(Δt) ∂t ∂x ∂x
f jn = f (t, x j ) ⎝ ∂t ⎠ j Δt
gives!
f jn+1 = f (t + Δt, x j )
⎛ ∂f ⎞
n
f n − f jn−1
f jn+1 = f (t, x j + h) = j +1 + O(h2 ) f jn+1 − fjn f jn+1 − fjn−1 fjn+1 − 2 f jn + f jn−1
⎝ ∂x ⎠ j 2h +U =D + O(Δt, h 2 )
Δt 2h h2
f jn−1 = f (t, x j − h) n
⎜⎛ ∂ f ⎞ = f j +1 − 2 f j + f j −1 + O(h2 )
2 n n n
⎝ ∂x 2 ⎠ j Solving for the new value and dropping the error terms yields!
h2
UΔt n DΔt
f jn+1 = f jn − ( f j +1 − f jn−1) + 2 ( f jn+1 − 2 f jn + f jn−1)
2h h

Finite Difference Approximations!
Thus, given f at one time (or time level), f at the

next time level is given by:!
UΔt n DΔt
f jn+1 = f jn − ( f − f jn−1) + 2 ( f jn+1 − 2 f jn + f jn−1)
2h j +1 h Example!
The value of every point at level n+1 is given
explicitly in terms of the values at the level n!
f jn+1
f jn−1 f jn f jn+1 Δt
h
h



EX2!
% one-dimensional advection-diffusion by the FTCS scheme!

n=21; nstep=100; length=2.0; h=length/(n-1); dt=0.05; D=0.05;!
f=zeros(n,1); y=zeros(n,1); ex=zeros(n,1); time=0.0;!
A short MATLAB program! for i=1:n, f(i)=0.5*sin(2*pi*h*(i-1)); end;
for m=1:nstep, m, time!
% initial conditions!
for i=1:n, ex(i)=exp(-4*pi*pi*D*time)*...!

0.5*sin(2*pi*(h*(i-1)-time)); end; % exact solution !
The evolution of a sine wave is followed as it hold off; plot(f,'linewidt',2); axis([1 n -2.0, 2.0]); % plot solution!
is advected and diffused. Two waves of the hold on; plot(ex,'r','linewidt',2);pause; % plot exact solution!
y=f; % store the solution!
infinite wave train are simulated in a domain of for i=2:n-1,!
length 2. To model the infinite train, periodic f(i)=y(i)-0.5*(dt/h)*(y(i+1)-y(i-1))+...!
boundary conditions are used. Compare the ! D*(dt/h^2)*(y(i+1)-2*y(i)+y(i-1)); % advect by centered differences!
end;!
numerical results with the exact solution.! f(n)=y(n)-0.5*(dt/h)*(y(2)-y(n-1))+...!
! D*(dt/h^2)*(y(2)-2*y(n)+y(n-1)); % do endpoints for!
f(1)=f(n); % periodic boundaries!
time=time+dt;!
end;!


Summary!
Evolution
for!
U=1;
Finite difference approximations by Taylor
D=0.05;
expansion!
k=1
Movie: adv1!
Approximating a partial differential

N=21
equation!
Δt=0.05

Exact!
Numerical!

http://users.wpi.edu/~gretar/me612.html! Outline!

Elementary! ! !Finite difference approximations!
! !The linear advection-diffusion equation!
Numerical Analysis:! ! !Matlab code!
! !Accuracy and error quantification!
Finite Difference ! !Stability!
! !Consistency!
Approximations-III! ! !Multidimensional problems!
! !Steady state!
Grétar Tryggvason!
Spring 2010!


Evolution
for!
U=1;

D=0.05;

k=1
Movie: adv1!
Accuracy!
N=21

Δt=0.05

Exact!
Numerical!


Accuracy!
It is clear that although the numerical solution is

qualitatively similar to the analytical solution, there Evolution m=11; time=0.50!
are significant quantitative differences.! for!
U=1;
2
The derivation of the numerical approximations for D=0.05;

1.5
the derivatives showed that the error depends on k=1

1
the size of h and Δt. !

0.5
N=21

First we test for different Δt. ! Δt=0.05
0
-0.5
Number of time steps= T/Δt

-1
Exact!
-1.5
Numerical! -2
2 4 6 8 10 12 14 16 18 20

Accuracy! Accuracy!
Repeat with m=21; time=0.50! Repeat with m=41; time=0.50!

a smaller a smaller
time-step! 2 time-step! 2
Δt=0.025
1.5
Δt=0.0125
1.5
1 1
N=21
N=21

0.5 0.5
0 0
-0.5 -0.5
-1 -1
Exact! Exact!
-1.5 -1.5
Numerical! -2
Numerical! -2
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20


Accuracy!
Very fine spatial

How accurate solution can resolution and a m=1001; time=0.50!
we obtain?! small time step!
2
U=1;

Take ! D=0.05;
1.5
Δt = 0.0005 ! k=1
1
and ! 0.5
N=200! N=200
0
Δt=0.0005

-0.5
-1
Exact!
-1.5
Numerical! -2
20 40 60 80 100 120 140 160 180 200


Accuracy!
Examine the spatial accuracy by taking a

Quantifying the Error! very small time step, Δt = 0.0005 and vary
the number of grid points, N, used to
Order of Accuracy! resolve the spatial direction. !
The grid size is h = L/N where L = 1 for our

case!

Accuracy! Accuracy!
Exact! Accuracy. Effect of!

time=0.50!
time=0.50! spatial resolution!
100
Numerical!
dt=0.0005!
N=11; E = 0.1633! N=11 to N=161! 10-1
N =21; E = 0.0403!
N =41; E = 0.0096!
-2
N =61; E = 0.0041! E
10
N =81; E = 0.0022!
N =101; E = 0.0015! 10-3
N N =121; E = 0.0011!
E=h ∑( f
j =1
j − fexact )2 N =161; E = 9.2600e-04! 10-4 0
10 10
1
10
2
1/ h


Accuracy! Accuracy!
If the error is of second order:! time=0.50!

−2 Accuracy. Effect of! 100
⎛1
E = Ch 2 = C ⎞ spatial resolution!
⎝h ⎠ 10-1
Taking the log:!

-2
E
10 2!
⎛ ⎛ 1 ⎞−2 ⎞ ⎛1
ln E = ln⎜ C ⎟ = lnC − 2ln ⎞ 1!
⎝ ⎝ h ⎠ ⎠ ⎝h ⎠ 10-3
dt=0.0005!
On a log-log plot, the E versus (1/h) curve 10-4 0 1 2
10 10 10
should therefore have a slope -2! N=11 to N=161!
1/ h


Accuracy! Summary!
Why is does the error deviate from the line for the
highest values of N?!
Showed, by numerical experiments, that
accuracy increases as the resolution is
Truncation! increased!
Error!
Showed that the error behaves in a way

that should be predictable!
Total!
Roundoff!
Number of Steps, N = 1/h!


http://users.wpi.edu/~gretar/me612.html! Outline!

! !Finite difference approximations!
Elementary! ! !The linear advection-diffusion equation!
! !Matlab code!
Numerical Analysis:! ! !Accuracy and error quantification!
! !Stability!
Finite Difference ! !Consistency!
! !Multidimensional problems!
Approximations-IV! ! !Steady state!
Grétar Tryggvason!
Spring 2010!


Stability! Stability!
Evolution
As long as accuracy is reasonable, integration at for!
larger time steps is more efficient and desirable. ! U=1;

D=0.05;

Can we increase the time step indefinitely?! k=1

Letʼs repeat the 1-D advection-diffusion equation N=21

with larges time step.! Δt=0.2

Use Δt = 0.2, instead of Δt = 0.05!
Exact!
Numerical!


Stability!
Instead of decaying as it should, the

amplitude of the numerical solution keeps
increasing. !
Ordinary Differential Equation!
Indeed, if we continued the calculations, we Stability!
would eventually produce numbers larger
than the computer can handle. This results
in an “overflow” or “NaN” (Not a Number).!

ODE—Example! ODE—Example!
5
Take:! f n+1 = (1 − Δt) f n 4

df
= −f with initial condition! f (0) = 1
3
dt Δt=0.5!
2
1
The exact solution is!
Δt=1.5! 0
f (t ) = e −t Δt=2.5! -1
-2
Forward Euler! -3
-4
f n+1
= f − f Δt
n n
f n+1
= (1 − Δt) f n
-5
0 5 10 15 20
n +1 n +1
f f
≤1 only if! Δt ≤ 2 ≤1 only if! Δt ≤ 2
fn fn


ODE—Example!
However!
f n +1 = (1− Δt) f n = (1− Δt) 2 f n−1
= ………… = (1− Δt) n f 1
Obviously, f oscillates unless! Δt ≤ 1 Stability Analysis of the!
Advection-Diffusion Equation:!
Backward Euler! Von Neumann Method!
f n +1 1
f n+1 = f n − f n+1 Δt =
fn (1+ Δt)
f n +1
≤1 for all Δt

fn


Consider the 1-D advection-diffusion equation:!

Generally, stability analysis of the full nonlinear system of
∂f ∂f ∂2f
equations is too involved to be practical, and we study a
+U =D 2
model problem that in some way mimics the full equations.! ∂t ∂x ∂x
The linear advection-diffusion equation is one such model
equation, and we will apply von Neumann's method to In finite-difference form:!
check the stability of a simple finite difference
approximation to that equation.!
f jn +1 − f jn f jn+1 − f jn−1 f jn+1 − 2 f jn + f jn−1
+U =D
Δt 2h h2
Look at the evolution of a small perturbation!
f jn = ε nj

The evolution of the perturbation is governed by:! The error at node j is:!
ikx j
n +1 ε nj = ε n e
ε j −ε n
j ε n
j +1 −ε n
j −1 ε n
j +1 − 2ε + ε
n
j
n
j −1
+U =D
Δt h2
2h The error at j+1 and j -1 can be written as!
ikx j +1 ik (x j + h) ikx j ikh
Write the error as a wave (expand as a Fourier series):! ε nj +1 = ε n e = εne = εne e
∞
ε nj = ε n ( x j ) = ∑ε
k = −∞
n ikx j
ke ε nj −1 = ε n e
ikx j −1 ik (x j −h)
= ε ne = ε ne
ikx j
e −ikh
Dropping the subscript!

ikx j
ε nj = ε n e
Recall:!
e ikx = cos kx + isin kx


Substituting! The equation for the error is:!

ε =ε e
n
j
n ikx j
ε n
j +1 =ε e
n ikx j
e
ikh
ε n
j −1 =ε en ikx j
e −ikh
ε n+1 − ε n ε n ikh − ikh ε n ikh −ikh
+U (e − e ) = D 2 (e − 2 + e )
into! Δt 2h h
ε n+1 − ε nj ε n − ε nj −1 ε n − 2ε nj + ε nj −1
+ U j +1 = D j +1
j
Δt 2h h2
yields! Solving for the ratio of the errors:!
ε n+1e − ε ne εn
ikx j ikx j
+ U (eikhe ikx j − e− ikheikx j ) =
Δt 2h ε n +1 UΔt ikh −ikh DΔt ikh
= 1− (e − e ) + 2 (e − 2 + e −ikh )
εn εn 2h h
D 2 (e ikheikx j − 2e ikx j + e − ikhe ikx j )
h


Dividing by the error amplitude at n:! The ratio of the error amplitude at n+1 and n is:!
amplification factor! ε n +1 DΔt h UΔt
= 1− 4 2 sin 2 k − i sin kh
ε n +1
UΔt ikh −ikh DΔt ikh εn h 2 h
= 1− (e − e ) + 2 (e − 2 + e −ikh )
εn 2h h Stability requires that!
UΔt DΔt ε n+1

≤1
= 1− isin kh + 2 2(cos kh −1) εn
h h
DΔt 2 h UΔt Since the amplification factor is a complex number, and k,
= 1− 4 2 sin k − i sin kh the wave number of the error, can be anything, the
h 2 h determination of the stability limit is slightly involved.!
Using:!
We will look at two special cases: (a) U = 0 and (b) D = 0
e ikh + e−ikh = 2cos kh; e ikh − e−ikh = i2sin kh; 2sin 2 θ = 1− cos2θ

(a) Consider first the case when U = 0, so the problem (b) Consider now the other limit where D = 0 and
reduces to a pure diffusion! we have a pure advection problem.!
ε n+1 DΔt h ε n +1 UΔt
= 1− 4 2 sin2 k = 1− i sin kh
εn h 2 εn h
2
Since sin () ≤ 1 the amplification factor is always ε n+1
Since the amplification
less than 1, and we find that it is bigger than -1 if!
factor has the form 1+i() εn
DΔt the absolute value of this
−1≤ 1− 4 ≤1
h2 complex number is always UΔt
larger than unity and the i sin kh
DΔt 1 method is unconditionally h
2 ≤
h 2 unstable for this case.!
1


For a two-dimensional problem, assume an error of the form!

i( kxi +ly j )
For the general case we must investigate the stability ε i,n j = ε n e
condition in more detail. We will not do so here, but
simply quote the results:! A stability analysis gives:!
ΔtD 1 U 2 Δt DΔt 1 (| U | + | V |) 2 Δt
≤ and! ≤2 ≤ and! ≤4
h2 2 D h2 4 D
Notice that high velocity and low viscosity lead to For a three-dimensional problem we get:!
instability according to the second restriction.!
DΔt 1 (|U | + |V | + |W |) 2 Δt
≤ and! ≤8
h2 6 D


Stability! Summary!
Stability – Now you know!!

Convergence – the solution to the finite-difference
equation approaches the true solution to the PDE having Introduced a formal method to examine
the same initial and boundary conditions as the mesh is whether a given finite difference
refined.! approximation is stable or not!
Laxʼs Equivalence Theorem! Introduced Laxʼs equivalent theorem.!

Given a properly posed initial value problem and a finite-
difference approximation to it that satisfies the consistency
condition, stability is the necessary and sufficient condition
for convergence.!

2010 Lecture 2

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

2010 Lecture 2

Hochgeladen von

Copyright:

Verfügbare Formate

Computational Fluid Dynamics I

Computational Fluid Dynamics I

Introduce the basic concepts needed to solve a partial

Numerical Analysis:! Discuss basic time integration methods, ordinary and

Approximations-II! Show the implementation of numerical algorithms into

Computational Fluid Dynamics I

• Solving partial differential equations!

Computational Fluid Dynamics I

Derive a numerical approximation to the governing

using a Taylor series !

Solving this equation for the time derivative gives:!

Computational Fluid Dynamics I

Apply this to the ODE from last lecture:!

The original equation at time level n is:!

Computational Fluid Dynamics I

For initial conditions of the form:!

to demonstrate how to solve a partial equation numerically.!

Although this equation is much simpler than the full

When using FINITE DIFFERENCE approximations,

The derivatives of the function are approximated

Computational Fluid Dynamics I

Start by expressing the value of f(x+h) and f(x-h) in terms of f(x)

Subtracting the second equation from the first:! ∂f (x) f (x + h) − f (x − h) ∂ 3 f (x) h 2

Computational Fluid Dynamics I

Start by expressing the value of f(x+h) and f(x-h) in terms of f(x)

∂f (x) ∂ 2 f (x) h 2 ∂ 3 f (x) h 3 ∂ 4 f (x) h 4

Adding the second equation to the first:!

The result is:!

Computational Fluid Dynamics I

We already introduced the notation! f = f (t) A numerical approximation to!

For space and time we will use:! ∂f ∂f ∂2 f

Computational Fluid Dynamics I

Using our shorthand ! Substituting these approximations into:!

Thus, given f at one time (or time level), f at the

Computational Fluid Dynamics I

% one-dimensional advection-diffusion by the FTCS scheme!

for i=1:n, ex(i)=exp(-4*pi*pi*D*time)*...!

Computational Fluid Dynamics I

Approximating a partial differential

• Solving partial differential equations!

Computational Fluid Dynamics I

Computational Fluid Dynamics I

It is clear that although the numerical solution is

The derivation of the numerical approximations for D=0.05;

the derivatives showed that the error depends on k=1

the size of h and Δt. !

Number of time steps= T/Δt

Repeat with m=21; time=0.50! Repeat with m=41; time=0.50!

Computational Fluid Dynamics I

Very fine spatial

Computational Fluid Dynamics I

Examine the spatial accuracy by taking a

The grid size is h = L/N where L = 1 for our

Exact! Accuracy. Effect of!

Computational Fluid Dynamics I

If the error is of second order:! time=0.50!

Taking the log:!

Computational Fluid Dynamics I

Showed that the error behaves in a way

Number of Steps, N = 1/h!

• Solving partial differential equations!

Computational Fluid Dynamics I

Computational Fluid Dynamics I

•  Solving partial differential equations!

for i=1:n, ex(i)=exp(-4pipiDtime)*...!

•  Solving partial differential equations!

•  Solving partial differential equations!